


What Makes A Friendly Face? Citizen Science to Create Sympathetic Android Faces

by Sorceringing



Category: Detroit: Become Human (Video Game)
Genre: Academic Paper, Androids, Gen, Research, Research Paper, how its made
Language: English
Status: Completed
Published: 2019-05-17
Updated: 2019-05-17
Packaged: 2020-03-06 22:43:05
Rating: General Audiences
Warnings: No Archive Warnings Apply
Chapters: 1
Words: 3,732
Publisher: archiveofourown.org
Story URL: https://archiveofourown.org/works/18860443
Author URL: https://archiveofourown.org/users/Sorceringing/pseuds/Sorceringing
Summary: Ever wondered how every android got their face? I doubt you'd want to see an existing person's appearance mass produced a billion times. So how did androids get their looks and what kind of research goes behind crafting that many aesthetically pleasing android faces so quickly?This peer-reviewed academic research paper discusses the work of scientist Nico Oh and associates as he presents his research on android aesthetics development and its benefits and implications.(basically it's a fake research paper set in the dbh universe about how this guy developed a model to create aesthetically pleasing faces with. You know how most people just write a post about their headcanon? Well this headcanon was taken wildly too far and now has irl sources)





	What Makes A Friendly Face? Citizen Science to Create Sympathetic Android Faces

**Author's Note:**

> WARNING. This paper is set in the DBH universe but uses REAL LIFE sources to back up the research statements. It also mixes images of FAKE faces with REAL ones and it can be difficult to determine the difference between the two. If you're sensitive to not knowing what's real and what isn't, please take care before reading. (notes at the end!)

# What Makes A Friendly Face?  
Citizen Science to Create Sympathetic Android Faces

Nico Oh1,2, Adrian Li1,2, Alexander Rebane1,3, Elijah Kamski4  
N.Oh@science.cyberlife.com  
1 Android Cosmetics, Research Division, CyberLife  
2Android-Human Relations, Research Division, CyberLife  
3Android AI, Research Division, CyberLife  
4Faculty of Science, CyberLife

**Abstract**  
With the creation and sale of the ST200 model, the demand for more diverse androids has become more prominent. In order to maximize the chance of producing aesthetically pleasing appearances, CyberLife scientists have teamed up with citizens all over the world in order to develop aesthetically pleasing faces for the next generation androids. A generative adversarial network using a style-based generator architecture was created to generate faces which were rated by the public on attractiveness in three categories: androids for homelife, work life or service. Every three months the model was updated with the most attractive faces in order to improve model output. The results of the study showed that the faces which were rated most attractive were often average in appearance rather than hypermasculine or hyperfeminine. Feminine faces were preferred with natural-looking make-up over no make-up, they were also younger looking than the masculine faces picked. Masculine faces were preferred clean shaven or with light stubble over beards and generally soft features and a good jawline were desirable. These results will be taken into account with the release of the latest models and are a good indicator of which traits make desirable faces for future products. Moreover, this study had an added benefit that the study fostered a relationship between consumers and potential products, making them more likely to purchase the finished result in the future.

 **Keywords:** Citizen Science, Machine Learning, Androids

 **1    Introduction**  
In 2022, the RT600 android known as “Chloe” became the first android to ever pass the Turing test live on television. “Chloe” demonstrated a sense of humanity during the test no android before it had ever managed to capture. This was partially due to it possessing the latest technologies developed by Elijah Kamski and colleagues, namely Thirium-310 and biocomponents[1]–[3]. The publicity gained through the televised test sent public demand for a “Chloe” of their own skyrocketing, which resulted in the development and sale of the ST200 model, affectionately known as “hostess Chloe”. A model that remains popular to this day [4].

Since that initial success, the public has expressed a desire for a more diverse line of androids to integrate into the workforce and home life. However, the development of the RT600’s face alone took several years and afterwards, only one face was produced [1]. This method would not be viable for the current fast paced market. In order to stay relevant, a quicker method for creating sympathetic faces must be developed.

One of the biggest limitations in the initial development of the RT600’s face, was the laborious process of designing and sculpting the face. The Chloe model’s gentle expression and smile were specifically designed to invoke trust in humans and is the result of months of focus group testing using artist sketches and both 3D and sculpted models. With the acquisition of NVIDIA’s Generative Adversarial Networks [5][6], CyberLife has acquired the ability to produce 2D images of human faces at a much higher speed, however sorting through the data using focus groups as done with Chloe would not be time nor cost effective. Developing an algorithm that can pinpoint the most statistically pleasant face would also not guarantee that the android will be well received upon release as people have varying opinions on what would make a face friendly or trustworthy. Instead, a novel approach to creating android faces has been developed involving the public’s opinion early on in the process using citizen science.

Citizen science is the practice where scientists recruit citizens as researchers rather than have them just as data collectors. Citizen scientists are given access to data to analyze themselves and their results are used in the ongoing research [7]. In the past, citizen science has already proven to be a valuable, low-cost method of performing large scale research, especially for environmental science and astronomy. An example of such success is the popular Zooniverse website [8], where users can log on and participate in a variety of citizen science projects ranging from arts to biology to physics. Within the Zooniverse website the astronomy projects Galaxy Zoo and Galaxy Zoo 2 are some of the more acclaimed citizen science projects where users view galaxy images and determine its shape. These projects have resulted in 58 publications as of 2019 and helped scientists to identify a variety of galaxy characteristics [9]. 

We propose that this method of technology-mediated citizen science will produce aesthetically pleasing android faces for future androids by allowing citizens into the process of shaping an android’s face early on. Involving the public at the start of the process may also increase the likelihood of a positive reception on the market after its release. During the project, citizens will experience the android’s entire creation process from idea to release. They may therefore form an affectionate bond with the product even before its release, making them more likely to purchase it [10].

This paper reports the findings of CyberLife’s Hello Detroit! citizen science project in which participants were asked to rate the generated faces on friendliness, attractiveness and trustworthiness amongst other things. The results of these findings were then used to optimize the face generator to produce more aesthetically pleasing faces until finally three optimized faces were picked for production. Implications and other findings will also be shortly discussed.

 **2    Methods**  
2.1    The face generator  
The face generator was built using a generative adversarial network using a style-based generator architecture [5] and initially used images from the Flickr-Faces-HQ (FFHQ) dataset. The generator started with two sources (source A and source B) as latent codes and based on them, different sets of images were generated and stored in a database to use in the Hello Detroit! project. Every 3 months, the latent codes were be updated with the faces that received the most positive feedback from the users and the generator generated new faces based on those sources. This continued for a total of 12 months, after which the top 6 well received faces were picked for production.

2.2    The rating system  
The rating system used in the project featured 5 Likert scales (1-5) and a comment box for manual input where they were asked to describe the features of the face they were viewing and add any other comments they may have had (Figure 1a). Users could use the Likert scale to indicate how much they agreed or disagreed with the five given statements. Of the five statements for each category, two questions were generic (naturalness of face and attractiveness) and three were specific to the task the android was geared towards.

Alongside the Likert scale questions, users were encouraged to write down their own feelings towards the faces using the text box provided. These textual inputs were processed into word networks.

2.3    The Mobile App  
The aforementioned set-up was also made available for mobile users accessing the CyberLife website on their mobile browser, however an app under the same name as the project was also provided. This app included a Tinder-like interface for rating android faces. Users could swipe right to indicate they found the face pleasing or left if they found the face in some way lacking (Figure 1b).

2.4    Statistics and normalisation  
As this study is of a qualitative nature, statistical significance was deemed less important than raw results, however normalisation techniques were applied in order to prevent bias towards active users. This was done by adding a weighting factor. Very active accounts would eventually gain less importance in the dataset.

2.5    Data collection  
Upon registration, each user registers their email address, name, age, and gender. If a social media profile was used, all publicly accessible fields with regards to identity was stored. Users could also opt to provide more information, such as: sexuality, ethnicity/race, education level, country of residence and economic status. The data collected from the user was used to examine biases and preferences within populations in the study and was also used by the users to compare their own results to the general results of the study anonymously. Two questionnaires were filled in at the start and at the end of the study to inquire about personal opinions such as stance on technology and towards AI and androids.

To preserve the privacy of the users, users were defined by IDs and no names were linked to any of the ratings or comments given during the study. Rather, name was only used to personalize on-site experiences and emails.

  
Figure 1. a) Standard layout of a Hello Detroit! rating page on a web browser. b) standard layout of a Hello Detroit! rating page on mobile.

 **3    Results**  
About 1 200 000 users registered an account on the website. Of these users, 46 per cent live within the United States of America, with around 20 per cent living in Detroit. The remaining percentages were divided between Europe and Asia. Most voters identified as straight, followed by bi and gay. Furthermore, slightly more men (age range 18 to 68) than women (age range 18 to 70) registered and a small fraction (4 per cent of the total population) identified as nonbinary or gender non-conforming. Voters who registered their ethnicity/race were predominantly European White, followed by Black, Asian and Hispanic.

The first batch of generated faces was created using the pre-existing FFHQ dataset (Figure 2). Most of the faces generated by this batch were rated as average in attractiveness and low in naturalness. Many users noted that the generated faces were slightly asymmetrical when the face contained make up. Others commented that completely symmetrical faces without blemishes were unnerving rather than pleasing. The most common descriptives, excluding the term "attractive", used for describing positively rated feminine faces were "pretty", "beautiful" and "cute", whereas positively rated masculine faces were described as "handsome", "good looking" and "striking". Displeasing feminine faces were commonly described as "ugly", "old" and "plain". Displeasing masculine faces were also often described as "ugly", along with "angry" and "creepy".

  
Figure 2. Example images from the Flickr-Faces-HQ dataset on which the first face generator was trained on

When examining the data per category, young faces were more likely to be considered attractive and desirable to be around in service or homelife, whereas on the workforce a variety of faces were considered positive, however young faces generally still rated highest. This also reflected on the quicker the mobile app variant of the study, voters swiped positively for young faces more often than for old faces when it came to service and homelife faces. Whereas for workforce faces, both young and older faces were considered appealing.

The second and third generation of faces yielded similar responses, however the naturalness of the face improved over time. Furthermore, the amount of displeasing faces decreased per generation and the average attractiveness for all faces improved significantly, especially in the third-generation model. More comments were left with remarks on how realistic the face looked and fewer negative comments were left.

By the fourth generation, the variety in feminine faces had decreased to younger faces of different ethnicities. The highest rated faces within this category contained light natural-looking make-up such as eyeliner, blush and lipstick or lip gloss. Masculine faces were young but remained relatively variable in facial features, a strong jawline was preferred over a soft one and clean shaven to slight stubble was preferred over bearded looks. Notably, positively rated feminine faces were generally young faces resembling people of early twenties to early thirties. Whereas positively rated masculine faces had a wider age range to them. Ambiguous faces were rated in a similar fashion as masculine faces. It was also found that the highest rated attractive masculine and feminine face were not hypermasculine nor hyperfeminine but rather looked average in appearance.

Interestingly, when sorting answers by gender, age and sexuality; straight male voters of all ages rated feminine faces higher in attractiveness than masculine faces on average. Older male voters (+50) did not often rate feminine faces of their age as attractive and often only assigned high attractive scores to young faces. Straight female voters found both masculine and feminine faces equally attractive and still voted older faces as attractive if they were around their age as well. Gay or otherwise not-straight voters of any gender on average rated faces similarly, with a slight skew towards their preference.

In the last month of the study the scores of the faces generated by the latest model were tallied and 6 faces were selected for production, two for each category. The service and homelife faces were notably softer than the faces for workforce only (figure 3). However, other high-ranking and low-ranking faces were stored in a separate database for potential future use (figure 4).

  
Figure 3. Six highest rated faces in no particular order to preserve confidentiality for release.

  
Figure 4. Other examples of faces in the fourth-generation network.

The results of the questionnaire were positive, users with a hesitant stance towards AI and android were found to have changed their mind after the study as they were now more educated about what an AI and android is and what it can and cannot do for society. On average, all participants were also more likely to feel comfortable with purchasing an android in the future after participating in the study.

 **4    Discussion and conclusion**  
4.1    Discussion of study results  
Ultimately, the highest rated masculine and feminine faces were selected for production (figure 3). Based on the data, younger faces were generally more desirable than older ones. Furthermore, faces without imperfections or blemishes were more often considered uncanny and unsettling rather than appealing. On feminine faces, light natural-looking make-up was preferred on the face rather than no make-up at all. On masculine faces, clean shaven to slight stubble was preferred over bearded looks and a defined jawline was more often met with higher attractiveness than faces without one. 

When examining the final two faces per category, its notable that none of the highest rated faces have hyperfeminine nor hypermasculine traits but resembled average everyday faces, which adheres to previous research on facial attraction [11]. Between the different generations of faces crafted by the model, there was an upwards trend in attractiveness, which is likely due to the fact that computer-generated average faces are often more attractive than the individual faces that make up their design [11]. Hence, the model, which kept being trained with faces which were rated with a high attractiveness kept producing more attractive faces over time. However, although the faces may be more attractive, there is also a threshold that can be crossed where the generated face becomes more unsettling rather than attractive. A phenomenon referred to as the uncanny valley [12]. This phenomenon is usually caused by abnormal features apart from the level of realism in the face, which may also explain why the most attractive generated faces were average in nature.

In terms of bias, there was a slight bias towards white American preferences as they made up most of the study population, however normalization techniques were applied to attempt to minimize this effect. In general, due to the number of participants, the chosen faces can be considered the average most attractive faces overall, as it was chosen based on global opinions. This also helps combat racial and gender bias. However, it is acknowledged that there will always be a number of people who will find the chosen faces unattractive or unappealing. Attractiveness in general is a subjective experience as individual preferences, among other factors, play a role in what a person may perceive as attractive [11].

To conclude, the GAN model was able to produce realistic and aesthetically pleasing faces within a year of its creation, enough so to publish on and to use in future studies. The faces generated will be stored for use in future iterations of the model and the opinions of the users will likely help create even more android faces in the future. Further research will involve finding which voice and face combo is the most pleasing by introducing soundbites of generated voices to each face and letting users rate the combo. Furthermore, this research as it is now can be extended into psychology in order to determine a global standard for pleasing humanoid aesthetics in other applications such as user interfaces, game design or mannequin design. It can also be used to study how personal biases affect perception on if a face is friendly or not.

4.2    General discussion  
Although other companies have practiced consumer engagement during product development before, this study marks one of the first instances where not only a new product was manufactured, but also scientific data was produced that examines global biases on what is considered an attractive persona. This will help other manufacturers create better designs for future products as mentioned before. It is also the first in its kind to use consumer engagement not for consumers to create their own personas in a video game, but to create an entirely separate persona that would likely enter their own homes and lives one day.

Furthermore, this method of involving citizens and other people into the process of designing a face is a cheaper and faster method of designing new products. By allowing people to sort through images to update the model with, less processing power is needed to define what makes a friendly face by the computer itself. This allows for the production of genuine, human-approved, friendly faces rather than perhaps statistically pleasing faces that end up unappealing towards humans. It is also faster than organizing focus groups for testing as every rater does this from the comfort of their own home, at their own pace and there is no need for organizing.

Another added benefit to this study was the unexpected popularity it received online. Pictures generated by the model were shared online as memes and other edits were made of them. This created unintentional media hype which spread all across the world allowing more people to participate in the study. The reach generated by this hype was far greater than any targeted or planned advertising method would have had as it spread through person-to-person niche channels not usually targeted by mainstream ad campaigns. Furthermore, it utilizes humor in a way not understood on a corporate level [13]. The Hello Detroit! subreddit (no affiliation), boasted about 100 000 users and was actively used and moderated by participants of the study. Users would post theories on what CyberLife would be using the faces they pick on other than just androids, talk about their favorite and least favorite faces and generally build a community for the study by discussing their own findings and comparing their research stats with others or discussing CyberLife emails with each other.

The formation of a community around the Hello Detroit! project likely also influenced how positively people responded at the end of the study. More people felt engaged and involved with the product and its manufacturing process. Which benefits the company and may increase sales of the new androids. Users reported feeling a more personal connection due to them being a part of the creation process and seemed more receptive to the idea of letting androids into their homes and workplaces.

All in all, this novel study showcased the efficacy of involving potential clients and citizens in general into the process of creating new products and into the field of scientific research. Researchers save time and money on their study and gain valuable insight on public opinions, whereas citizens gain insight on behind-the-scenes practices in research. This further proves that citizen science is a very beneficial tool within science and is not limited to biological sciences only. It should also not be disregarded within the field of technology research [14].

 **Acknowledgements**  
Many thanks to my supervisor and the director of CyberLife, Elijah Kamski for providing the funds and office space used during the project.

Special thanks to Alexander Rebane for help on designing the Hello Detroit! webpages and Adrian Li for the in-depth discussions on content for this paper. I owe you guys dinner and a coffee next time we leave the basement.

 **References**  
[1] E. Kamski and A. Stern, “Chloe: An autonomous general-purpose domestic companion,” Sci. Robot., vol. 5, no. 2, pp. 223–226, 2022.  
[2] E. Kamski, M. M. Fairchild, P. D. Raines, and Y. J. Ruan, “Thirium-310 for enhancing information and energy transfer in androids,” Fuel, vol. 2, no. 4, pp. 149–159, 2019.  
[3] E. Kamski, W. J. Slováček, and N. Kolijn, “Biotechnological components for simulating idle human behavior in androids,” Sci. Robot., vol. 9, no. 6, pp. 58–75, 2020.  
[4] L. Howard, “One million Chloes,” Business Insider, pp. 13–14, 2023.  
[5] T. Karras and T. Aila, “A Style-Based Generator Architecture for Generative Adversarial Networks.”  
[6] T. Karras, T. Aila, S. Laine, and J. Lehtinen, “Progressive growing of GANs for improved quality, stability, and variation,” in ICLR 2018, 2018, pp. 1–26.  
[7] R. Bonney et al., “Citizen Science : A Developing Tool for Expanding Science Knowledge and Scientific Literacy,” vol. 59, no. 11, pp. 977–984, 2009.  
[8] “About — Zooniverse.” [Online]. Available: https://www.zooniverse.org/about. [Accessed: 13-Apr-2019].  
[9] “Zooniverse - Publications.” [Online]. Available: https://www.zooniverse.org/about/publications.  
[10] J. Fuller, “Who Made This Thing? How Designer Identity and Brand Personality Impact Consumers’ Evaluations of New Product Offerings,” Adv. Consum. Res., vol. 33, pp. 212–217, 2006.  
[11] B. Fink and I. Penton-Voak, “Evolutionary psychology of facial attractiveness,” Curr. Dir. Psychol. Sci., vol. 11, no. 5, pp. 154–158, 2002.  
[12] J. Seyama and R. S. Nagayama, “The uncanny valley: Effect of realism on the impression of artificial human faces,” Presence Teleoperators Virtual Environ., vol. 16, no. 4, pp. 337–351, 2007.  
[13] C. linxia, “Memes in Advertising Slogans,” 2006. [Online]. Available: http://en.cnki.com.cn/Article_en/CJFDTOTAL-TEAC200604010.htm. [Accessed: 16-May-2019].  
[14] D. Brossard, B. Lewenstein, and R. Bonney, “Scientific knowledge and attitude change: The impact of a citizen science project,” Int. J. Sci. Educ., vol. 27, no. 9, pp. 1099–1121, 2005.  


**Author's Note:**

> All sources in the reference section except the kamski papers are real sources you can look up online! I highly recommend checking out the Zooniverse citizen science website as it features a lot of fun projects to do at your leisure if you're into animal watching or galaxies. They're a fun thing to pass time with!
> 
> The faces in Figure 4 (Apart from the faces of the actors in DBH of course) were generated using NVIDIA's GAN generator which you can find at thispersondoesnotexist.com . Check it out! the technology used in this paper already exist irl!
> 
> I'd also like to do a huge shout out to my friends for helping me finish this paper and also for putting up with all my complaining. <3 love you guys.


End file.
